Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * net/sched/sch_generic.c Generic packet scheduler routines.
4 *
5 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
7 * - Ingress support
8 */
9
10#include <linux/bitops.h>
11#include <linux/module.h>
12#include <linux/types.h>
13#include <linux/kernel.h>
14#include <linux/sched.h>
15#include <linux/string.h>
16#include <linux/errno.h>
17#include <linux/netdevice.h>
18#include <linux/skbuff.h>
19#include <linux/rtnetlink.h>
20#include <linux/init.h>
21#include <linux/rcupdate.h>
22#include <linux/list.h>
23#include <linux/slab.h>
24#include <linux/if_vlan.h>
25#include <linux/skb_array.h>
26#include <linux/if_macvlan.h>
27#include <net/sch_generic.h>
28#include <net/pkt_sched.h>
29#include <net/dst.h>
30#include <trace/events/qdisc.h>
31#include <trace/events/net.h>
32#include <net/xfrm.h>
33
34/* Qdisc to use by default */
35const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
36EXPORT_SYMBOL(default_qdisc_ops);
37
38/* Main transmission queue. */
39
40/* Modifications to data participating in scheduling must be protected with
41 * qdisc_lock(qdisc) spinlock.
42 *
43 * The idea is the following:
44 * - enqueue, dequeue are serialized via qdisc root lock
45 * - ingress filtering is also serialized via qdisc root lock
46 * - updates to tree and tree walking are only done under the rtnl mutex.
47 */
48
49#define SKB_XOFF_MAGIC ((struct sk_buff *)1UL)
50
51static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
52{
53 const struct netdev_queue *txq = q->dev_queue;
54 spinlock_t *lock = NULL;
55 struct sk_buff *skb;
56
57 if (q->flags & TCQ_F_NOLOCK) {
58 lock = qdisc_lock(q);
59 spin_lock(lock);
60 }
61
62 skb = skb_peek(&q->skb_bad_txq);
63 if (skb) {
64 /* check the reason of requeuing without tx lock first */
65 txq = skb_get_tx_queue(txq->dev, skb);
66 if (!netif_xmit_frozen_or_stopped(txq)) {
67 skb = __skb_dequeue(&q->skb_bad_txq);
68 if (qdisc_is_percpu_stats(q)) {
69 qdisc_qstats_cpu_backlog_dec(q, skb);
70 qdisc_qstats_cpu_qlen_dec(q);
71 } else {
72 qdisc_qstats_backlog_dec(q, skb);
73 q->q.qlen--;
74 }
75 } else {
76 skb = SKB_XOFF_MAGIC;
77 }
78 }
79
80 if (lock)
81 spin_unlock(lock);
82
83 return skb;
84}
85
86static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
87{
88 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
89
90 if (unlikely(skb))
91 skb = __skb_dequeue_bad_txq(q);
92
93 return skb;
94}
95
96static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
97 struct sk_buff *skb)
98{
99 spinlock_t *lock = NULL;
100
101 if (q->flags & TCQ_F_NOLOCK) {
102 lock = qdisc_lock(q);
103 spin_lock(lock);
104 }
105
106 __skb_queue_tail(&q->skb_bad_txq, skb);
107
108 if (qdisc_is_percpu_stats(q)) {
109 qdisc_qstats_cpu_backlog_inc(q, skb);
110 qdisc_qstats_cpu_qlen_inc(q);
111 } else {
112 qdisc_qstats_backlog_inc(q, skb);
113 q->q.qlen++;
114 }
115
116 if (lock)
117 spin_unlock(lock);
118}
119
120static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
121{
122 spinlock_t *lock = NULL;
123
124 if (q->flags & TCQ_F_NOLOCK) {
125 lock = qdisc_lock(q);
126 spin_lock(lock);
127 }
128
129 while (skb) {
130 struct sk_buff *next = skb->next;
131
132 __skb_queue_tail(&q->gso_skb, skb);
133
134 /* it's still part of the queue */
135 if (qdisc_is_percpu_stats(q)) {
136 qdisc_qstats_cpu_requeues_inc(q);
137 qdisc_qstats_cpu_backlog_inc(q, skb);
138 qdisc_qstats_cpu_qlen_inc(q);
139 } else {
140 q->qstats.requeues++;
141 qdisc_qstats_backlog_inc(q, skb);
142 q->q.qlen++;
143 }
144
145 skb = next;
146 }
147 if (lock)
148 spin_unlock(lock);
149 __netif_schedule(q);
150}
151
152static void try_bulk_dequeue_skb(struct Qdisc *q,
153 struct sk_buff *skb,
154 const struct netdev_queue *txq,
155 int *packets)
156{
157 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
158
159 while (bytelimit > 0) {
160 struct sk_buff *nskb = q->dequeue(q);
161
162 if (!nskb)
163 break;
164
165 bytelimit -= nskb->len; /* covers GSO len */
166 skb->next = nskb;
167 skb = nskb;
168 (*packets)++; /* GSO counts as one pkt */
169 }
170 skb_mark_not_on_list(skb);
171}
172
173/* This variant of try_bulk_dequeue_skb() makes sure
174 * all skbs in the chain are for the same txq
175 */
176static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
177 struct sk_buff *skb,
178 int *packets)
179{
180 int mapping = skb_get_queue_mapping(skb);
181 struct sk_buff *nskb;
182 int cnt = 0;
183
184 do {
185 nskb = q->dequeue(q);
186 if (!nskb)
187 break;
188 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
189 qdisc_enqueue_skb_bad_txq(q, nskb);
190 break;
191 }
192 skb->next = nskb;
193 skb = nskb;
194 } while (++cnt < 8);
195 (*packets) += cnt;
196 skb_mark_not_on_list(skb);
197}
198
199/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
200 * A requeued skb (via q->gso_skb) can also be a SKB list.
201 */
202static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
203 int *packets)
204{
205 const struct netdev_queue *txq = q->dev_queue;
206 struct sk_buff *skb = NULL;
207
208 *packets = 1;
209 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
210 spinlock_t *lock = NULL;
211
212 if (q->flags & TCQ_F_NOLOCK) {
213 lock = qdisc_lock(q);
214 spin_lock(lock);
215 }
216
217 skb = skb_peek(&q->gso_skb);
218
219 /* skb may be null if another cpu pulls gso_skb off in between
220 * empty check and lock.
221 */
222 if (!skb) {
223 if (lock)
224 spin_unlock(lock);
225 goto validate;
226 }
227
228 /* skb in gso_skb were already validated */
229 *validate = false;
230 if (xfrm_offload(skb))
231 *validate = true;
232 /* check the reason of requeuing without tx lock first */
233 txq = skb_get_tx_queue(txq->dev, skb);
234 if (!netif_xmit_frozen_or_stopped(txq)) {
235 skb = __skb_dequeue(&q->gso_skb);
236 if (qdisc_is_percpu_stats(q)) {
237 qdisc_qstats_cpu_backlog_dec(q, skb);
238 qdisc_qstats_cpu_qlen_dec(q);
239 } else {
240 qdisc_qstats_backlog_dec(q, skb);
241 q->q.qlen--;
242 }
243 } else {
244 skb = NULL;
245 }
246 if (lock)
247 spin_unlock(lock);
248 goto trace;
249 }
250validate:
251 *validate = true;
252
253 if ((q->flags & TCQ_F_ONETXQUEUE) &&
254 netif_xmit_frozen_or_stopped(txq))
255 return skb;
256
257 skb = qdisc_dequeue_skb_bad_txq(q);
258 if (unlikely(skb)) {
259 if (skb == SKB_XOFF_MAGIC)
260 return NULL;
261 goto bulk;
262 }
263 skb = q->dequeue(q);
264 if (skb) {
265bulk:
266 if (qdisc_may_bulk(q))
267 try_bulk_dequeue_skb(q, skb, txq, packets);
268 else
269 try_bulk_dequeue_skb_slow(q, skb, packets);
270 }
271trace:
272 trace_qdisc_dequeue(q, txq, *packets, skb);
273 return skb;
274}
275
276/*
277 * Transmit possibly several skbs, and handle the return status as
278 * required. Owning running seqcount bit guarantees that
279 * only one CPU can execute this function.
280 *
281 * Returns to the caller:
282 * false - hardware queue frozen backoff
283 * true - feel free to send more pkts
284 */
285bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
286 struct net_device *dev, struct netdev_queue *txq,
287 spinlock_t *root_lock, bool validate)
288{
289 int ret = NETDEV_TX_BUSY;
290 bool again = false;
291
292 /* And release qdisc */
293 if (root_lock)
294 spin_unlock(root_lock);
295
296 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
297 if (validate)
298 skb = validate_xmit_skb_list(skb, dev, &again);
299
300#ifdef CONFIG_XFRM_OFFLOAD
301 if (unlikely(again)) {
302 if (root_lock)
303 spin_lock(root_lock);
304
305 dev_requeue_skb(skb, q);
306 return false;
307 }
308#endif
309
310 if (likely(skb)) {
311 HARD_TX_LOCK(dev, txq, smp_processor_id());
312 if (!netif_xmit_frozen_or_stopped(txq))
313 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
314
315 HARD_TX_UNLOCK(dev, txq);
316 } else {
317 if (root_lock)
318 spin_lock(root_lock);
319 return true;
320 }
321
322 if (root_lock)
323 spin_lock(root_lock);
324
325 if (!dev_xmit_complete(ret)) {
326 /* Driver returned NETDEV_TX_BUSY - requeue skb */
327 if (unlikely(ret != NETDEV_TX_BUSY))
328 net_warn_ratelimited("BUG %s code %d qlen %d\n",
329 dev->name, ret, q->q.qlen);
330
331 dev_requeue_skb(skb, q);
332 return false;
333 }
334
335 return true;
336}
337
338/*
339 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
340 *
341 * running seqcount guarantees only one CPU can process
342 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
343 * this queue.
344 *
345 * netif_tx_lock serializes accesses to device driver.
346 *
347 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
348 * if one is grabbed, another must be free.
349 *
350 * Note, that this procedure can be called by a watchdog timer
351 *
352 * Returns to the caller:
353 * 0 - queue is empty or throttled.
354 * >0 - queue is not empty.
355 *
356 */
357static inline bool qdisc_restart(struct Qdisc *q, int *packets)
358{
359 spinlock_t *root_lock = NULL;
360 struct netdev_queue *txq;
361 struct net_device *dev;
362 struct sk_buff *skb;
363 bool validate;
364
365 /* Dequeue packet */
366 skb = dequeue_skb(q, &validate, packets);
367 if (unlikely(!skb))
368 return false;
369
370 if (!(q->flags & TCQ_F_NOLOCK))
371 root_lock = qdisc_lock(q);
372
373 dev = qdisc_dev(q);
374 txq = skb_get_tx_queue(dev, skb);
375
376 return sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
377}
378
379void __qdisc_run(struct Qdisc *q)
380{
381 int quota = dev_tx_weight;
382 int packets;
383
384 while (qdisc_restart(q, &packets)) {
385 /*
386 * Ordered by possible occurrence: Postpone processing if
387 * 1. we've exceeded packet quota
388 * 2. another process needs the CPU;
389 */
390 quota -= packets;
391 if (quota <= 0 || need_resched()) {
392 __netif_schedule(q);
393 break;
394 }
395 }
396}
397
398unsigned long dev_trans_start(struct net_device *dev)
399{
400 unsigned long val, res;
401 unsigned int i;
402
403 if (is_vlan_dev(dev))
404 dev = vlan_dev_real_dev(dev);
405 else if (netif_is_macvlan(dev))
406 dev = macvlan_dev_real_dev(dev);
407 res = netdev_get_tx_queue(dev, 0)->trans_start;
408 for (i = 1; i < dev->num_tx_queues; i++) {
409 val = netdev_get_tx_queue(dev, i)->trans_start;
410 if (val && time_after(val, res))
411 res = val;
412 }
413
414 return res;
415}
416EXPORT_SYMBOL(dev_trans_start);
417
418static void dev_watchdog(struct timer_list *t)
419{
420 struct net_device *dev = from_timer(dev, t, watchdog_timer);
421
422 netif_tx_lock(dev);
423 if (!qdisc_tx_is_noop(dev)) {
424 if (netif_device_present(dev) &&
425 netif_running(dev) &&
426 netif_carrier_ok(dev)) {
427 int some_queue_timedout = 0;
428 unsigned int i;
429 unsigned long trans_start;
430
431 for (i = 0; i < dev->num_tx_queues; i++) {
432 struct netdev_queue *txq;
433
434 txq = netdev_get_tx_queue(dev, i);
435 trans_start = txq->trans_start;
436 if (netif_xmit_stopped(txq) &&
437 time_after(jiffies, (trans_start +
438 dev->watchdog_timeo))) {
439 some_queue_timedout = 1;
440 txq->trans_timeout++;
441 break;
442 }
443 }
444
445 if (some_queue_timedout) {
446 trace_net_dev_xmit_timeout(dev, i);
447 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
448 dev->name, netdev_drivername(dev), i);
449 dev->netdev_ops->ndo_tx_timeout(dev);
450 }
451 if (!mod_timer(&dev->watchdog_timer,
452 round_jiffies(jiffies +
453 dev->watchdog_timeo)))
454 dev_hold(dev);
455 }
456 }
457 netif_tx_unlock(dev);
458
459 dev_put(dev);
460}
461
462void __netdev_watchdog_up(struct net_device *dev)
463{
464 if (dev->netdev_ops->ndo_tx_timeout) {
465 if (dev->watchdog_timeo <= 0)
466 dev->watchdog_timeo = 5*HZ;
467 if (!mod_timer(&dev->watchdog_timer,
468 round_jiffies(jiffies + dev->watchdog_timeo)))
469 dev_hold(dev);
470 }
471}
472
473static void dev_watchdog_up(struct net_device *dev)
474{
475 __netdev_watchdog_up(dev);
476}
477
478static void dev_watchdog_down(struct net_device *dev)
479{
480 netif_tx_lock_bh(dev);
481 if (del_timer(&dev->watchdog_timer))
482 dev_put(dev);
483 netif_tx_unlock_bh(dev);
484}
485
486/**
487 * netif_carrier_on - set carrier
488 * @dev: network device
489 *
490 * Device has detected acquisition of carrier.
491 */
492void netif_carrier_on(struct net_device *dev)
493{
494 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
495 if (dev->reg_state == NETREG_UNINITIALIZED)
496 return;
497 atomic_inc(&dev->carrier_up_count);
498 linkwatch_fire_event(dev);
499 if (netif_running(dev))
500 __netdev_watchdog_up(dev);
501 }
502}
503EXPORT_SYMBOL(netif_carrier_on);
504
505/**
506 * netif_carrier_off - clear carrier
507 * @dev: network device
508 *
509 * Device has detected loss of carrier.
510 */
511void netif_carrier_off(struct net_device *dev)
512{
513 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
514 if (dev->reg_state == NETREG_UNINITIALIZED)
515 return;
516 atomic_inc(&dev->carrier_down_count);
517 linkwatch_fire_event(dev);
518 }
519}
520EXPORT_SYMBOL(netif_carrier_off);
521
522/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
523 under all circumstances. It is difficult to invent anything faster or
524 cheaper.
525 */
526
527static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
528 struct sk_buff **to_free)
529{
530 __qdisc_drop(skb, to_free);
531 return NET_XMIT_CN;
532}
533
534static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
535{
536 return NULL;
537}
538
539struct Qdisc_ops noop_qdisc_ops __read_mostly = {
540 .id = "noop",
541 .priv_size = 0,
542 .enqueue = noop_enqueue,
543 .dequeue = noop_dequeue,
544 .peek = noop_dequeue,
545 .owner = THIS_MODULE,
546};
547
548static struct netdev_queue noop_netdev_queue = {
549 RCU_POINTER_INITIALIZER(qdisc, &noop_qdisc),
550 .qdisc_sleeping = &noop_qdisc,
551};
552
553struct Qdisc noop_qdisc = {
554 .enqueue = noop_enqueue,
555 .dequeue = noop_dequeue,
556 .flags = TCQ_F_BUILTIN,
557 .ops = &noop_qdisc_ops,
558 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
559 .dev_queue = &noop_netdev_queue,
560 .running = SEQCNT_ZERO(noop_qdisc.running),
561 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
562 .gso_skb = {
563 .next = (struct sk_buff *)&noop_qdisc.gso_skb,
564 .prev = (struct sk_buff *)&noop_qdisc.gso_skb,
565 .qlen = 0,
566 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.gso_skb.lock),
567 },
568 .skb_bad_txq = {
569 .next = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
570 .prev = (struct sk_buff *)&noop_qdisc.skb_bad_txq,
571 .qlen = 0,
572 .lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.skb_bad_txq.lock),
573 },
574};
575EXPORT_SYMBOL(noop_qdisc);
576
577static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
578 struct netlink_ext_ack *extack)
579{
580 /* register_qdisc() assigns a default of noop_enqueue if unset,
581 * but __dev_queue_xmit() treats noqueue only as such
582 * if this is NULL - so clear it here. */
583 qdisc->enqueue = NULL;
584 return 0;
585}
586
587struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
588 .id = "noqueue",
589 .priv_size = 0,
590 .init = noqueue_init,
591 .enqueue = noop_enqueue,
592 .dequeue = noop_dequeue,
593 .peek = noop_dequeue,
594 .owner = THIS_MODULE,
595};
596
597static const u8 prio2band[TC_PRIO_MAX + 1] = {
598 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
599};
600
601/* 3-band FIFO queue: old style, but should be a bit faster than
602 generic prio+fifo combination.
603 */
604
605#define PFIFO_FAST_BANDS 3
606
607/*
608 * Private data for a pfifo_fast scheduler containing:
609 * - rings for priority bands
610 */
611struct pfifo_fast_priv {
612 struct skb_array q[PFIFO_FAST_BANDS];
613};
614
615static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
616 int band)
617{
618 return &priv->q[band];
619}
620
621static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
622 struct sk_buff **to_free)
623{
624 int band = prio2band[skb->priority & TC_PRIO_MAX];
625 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
626 struct skb_array *q = band2list(priv, band);
627 unsigned int pkt_len = qdisc_pkt_len(skb);
628 int err;
629
630 err = skb_array_produce(q, skb);
631
632 if (unlikely(err)) {
633 if (qdisc_is_percpu_stats(qdisc))
634 return qdisc_drop_cpu(skb, qdisc, to_free);
635 else
636 return qdisc_drop(skb, qdisc, to_free);
637 }
638
639 qdisc_update_stats_at_enqueue(qdisc, pkt_len);
640 return NET_XMIT_SUCCESS;
641}
642
643static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
644{
645 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
646 struct sk_buff *skb = NULL;
647 int band;
648
649 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
650 struct skb_array *q = band2list(priv, band);
651
652 if (__skb_array_empty(q))
653 continue;
654
655 skb = __skb_array_consume(q);
656 }
657 if (likely(skb)) {
658 qdisc_update_stats_at_dequeue(qdisc, skb);
659 } else {
660 qdisc->empty = true;
661 }
662
663 return skb;
664}
665
666static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
667{
668 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
669 struct sk_buff *skb = NULL;
670 int band;
671
672 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
673 struct skb_array *q = band2list(priv, band);
674
675 skb = __skb_array_peek(q);
676 }
677
678 return skb;
679}
680
681static void pfifo_fast_reset(struct Qdisc *qdisc)
682{
683 int i, band;
684 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
685
686 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
687 struct skb_array *q = band2list(priv, band);
688 struct sk_buff *skb;
689
690 /* NULL ring is possible if destroy path is due to a failed
691 * skb_array_init() in pfifo_fast_init() case.
692 */
693 if (!q->ring.queue)
694 continue;
695
696 while ((skb = __skb_array_consume(q)) != NULL)
697 kfree_skb(skb);
698 }
699
700 if (qdisc_is_percpu_stats(qdisc)) {
701 for_each_possible_cpu(i) {
702 struct gnet_stats_queue *q;
703
704 q = per_cpu_ptr(qdisc->cpu_qstats, i);
705 q->backlog = 0;
706 q->qlen = 0;
707 }
708 }
709}
710
711static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
712{
713 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
714
715 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
716 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
717 goto nla_put_failure;
718 return skb->len;
719
720nla_put_failure:
721 return -1;
722}
723
724static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
725 struct netlink_ext_ack *extack)
726{
727 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
728 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
729 int prio;
730
731 /* guard against zero length rings */
732 if (!qlen)
733 return -EINVAL;
734
735 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
736 struct skb_array *q = band2list(priv, prio);
737 int err;
738
739 err = skb_array_init(q, qlen, GFP_KERNEL);
740 if (err)
741 return -ENOMEM;
742 }
743
744 /* Can by-pass the queue discipline */
745 qdisc->flags |= TCQ_F_CAN_BYPASS;
746 return 0;
747}
748
749static void pfifo_fast_destroy(struct Qdisc *sch)
750{
751 struct pfifo_fast_priv *priv = qdisc_priv(sch);
752 int prio;
753
754 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
755 struct skb_array *q = band2list(priv, prio);
756
757 /* NULL ring is possible if destroy path is due to a failed
758 * skb_array_init() in pfifo_fast_init() case.
759 */
760 if (!q->ring.queue)
761 continue;
762 /* Destroy ring but no need to kfree_skb because a call to
763 * pfifo_fast_reset() has already done that work.
764 */
765 ptr_ring_cleanup(&q->ring, NULL);
766 }
767}
768
769static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
770 unsigned int new_len)
771{
772 struct pfifo_fast_priv *priv = qdisc_priv(sch);
773 struct skb_array *bands[PFIFO_FAST_BANDS];
774 int prio;
775
776 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
777 struct skb_array *q = band2list(priv, prio);
778
779 bands[prio] = q;
780 }
781
782 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
783 GFP_KERNEL);
784}
785
786struct Qdisc_ops pfifo_fast_ops __read_mostly = {
787 .id = "pfifo_fast",
788 .priv_size = sizeof(struct pfifo_fast_priv),
789 .enqueue = pfifo_fast_enqueue,
790 .dequeue = pfifo_fast_dequeue,
791 .peek = pfifo_fast_peek,
792 .init = pfifo_fast_init,
793 .destroy = pfifo_fast_destroy,
794 .reset = pfifo_fast_reset,
795 .dump = pfifo_fast_dump,
796 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
797 .owner = THIS_MODULE,
798 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
799};
800EXPORT_SYMBOL(pfifo_fast_ops);
801
802struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
803 const struct Qdisc_ops *ops,
804 struct netlink_ext_ack *extack)
805{
806 void *p;
807 struct Qdisc *sch;
808 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
809 int err = -ENOBUFS;
810 struct net_device *dev;
811
812 if (!dev_queue) {
813 NL_SET_ERR_MSG(extack, "No device queue given");
814 err = -EINVAL;
815 goto errout;
816 }
817
818 dev = dev_queue->dev;
819 p = kzalloc_node(size, GFP_KERNEL,
820 netdev_queue_numa_node_read(dev_queue));
821
822 if (!p)
823 goto errout;
824 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
825 /* if we got non aligned memory, ask more and do alignment ourself */
826 if (sch != p) {
827 kfree(p);
828 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
829 netdev_queue_numa_node_read(dev_queue));
830 if (!p)
831 goto errout;
832 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
833 sch->padded = (char *) sch - (char *) p;
834 }
835 __skb_queue_head_init(&sch->gso_skb);
836 __skb_queue_head_init(&sch->skb_bad_txq);
837 qdisc_skb_head_init(&sch->q);
838 spin_lock_init(&sch->q.lock);
839
840 if (ops->static_flags & TCQ_F_CPUSTATS) {
841 sch->cpu_bstats =
842 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
843 if (!sch->cpu_bstats)
844 goto errout1;
845
846 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
847 if (!sch->cpu_qstats) {
848 free_percpu(sch->cpu_bstats);
849 goto errout1;
850 }
851 }
852
853 spin_lock_init(&sch->busylock);
854 /* seqlock has the same scope of busylock, for NOLOCK qdisc */
855 spin_lock_init(&sch->seqlock);
856 seqcount_init(&sch->running);
857
858 sch->ops = ops;
859 sch->flags = ops->static_flags;
860 sch->enqueue = ops->enqueue;
861 sch->dequeue = ops->dequeue;
862 sch->dev_queue = dev_queue;
863 sch->empty = true;
864 dev_hold(dev);
865 refcount_set(&sch->refcnt, 1);
866
867 if (sch != &noop_qdisc) {
868 lockdep_set_class(&sch->busylock, &dev->qdisc_tx_busylock_key);
869 lockdep_set_class(&sch->seqlock, &dev->qdisc_tx_busylock_key);
870 lockdep_set_class(&sch->running, &dev->qdisc_running_key);
871 }
872
873 return sch;
874errout1:
875 kfree(p);
876errout:
877 return ERR_PTR(err);
878}
879
880struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
881 const struct Qdisc_ops *ops,
882 unsigned int parentid,
883 struct netlink_ext_ack *extack)
884{
885 struct Qdisc *sch;
886
887 if (!try_module_get(ops->owner)) {
888 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
889 return NULL;
890 }
891
892 sch = qdisc_alloc(dev_queue, ops, extack);
893 if (IS_ERR(sch)) {
894 module_put(ops->owner);
895 return NULL;
896 }
897 sch->parent = parentid;
898
899 if (!ops->init || ops->init(sch, NULL, extack) == 0)
900 return sch;
901
902 qdisc_put(sch);
903 return NULL;
904}
905EXPORT_SYMBOL(qdisc_create_dflt);
906
907/* Under qdisc_lock(qdisc) and BH! */
908
909void qdisc_reset(struct Qdisc *qdisc)
910{
911 const struct Qdisc_ops *ops = qdisc->ops;
912 struct sk_buff *skb, *tmp;
913
914 if (ops->reset)
915 ops->reset(qdisc);
916
917 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
918 __skb_unlink(skb, &qdisc->gso_skb);
919 kfree_skb_list(skb);
920 }
921
922 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
923 __skb_unlink(skb, &qdisc->skb_bad_txq);
924 kfree_skb_list(skb);
925 }
926
927 qdisc->q.qlen = 0;
928 qdisc->qstats.backlog = 0;
929}
930EXPORT_SYMBOL(qdisc_reset);
931
932void qdisc_free(struct Qdisc *qdisc)
933{
934 if (qdisc_is_percpu_stats(qdisc)) {
935 free_percpu(qdisc->cpu_bstats);
936 free_percpu(qdisc->cpu_qstats);
937 }
938
939 kfree((char *) qdisc - qdisc->padded);
940}
941
942static void qdisc_free_cb(struct rcu_head *head)
943{
944 struct Qdisc *q = container_of(head, struct Qdisc, rcu);
945
946 qdisc_free(q);
947}
948
949static void qdisc_destroy(struct Qdisc *qdisc)
950{
951 const struct Qdisc_ops *ops = qdisc->ops;
952 struct sk_buff *skb, *tmp;
953
954#ifdef CONFIG_NET_SCHED
955 qdisc_hash_del(qdisc);
956
957 qdisc_put_stab(rtnl_dereference(qdisc->stab));
958#endif
959 gen_kill_estimator(&qdisc->rate_est);
960 if (ops->reset)
961 ops->reset(qdisc);
962 if (ops->destroy)
963 ops->destroy(qdisc);
964
965 module_put(ops->owner);
966 dev_put(qdisc_dev(qdisc));
967
968 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
969 __skb_unlink(skb, &qdisc->gso_skb);
970 kfree_skb_list(skb);
971 }
972
973 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
974 __skb_unlink(skb, &qdisc->skb_bad_txq);
975 kfree_skb_list(skb);
976 }
977
978 call_rcu(&qdisc->rcu, qdisc_free_cb);
979}
980
981void qdisc_put(struct Qdisc *qdisc)
982{
983 if (!qdisc)
984 return;
985
986 if (qdisc->flags & TCQ_F_BUILTIN ||
987 !refcount_dec_and_test(&qdisc->refcnt))
988 return;
989
990 qdisc_destroy(qdisc);
991}
992EXPORT_SYMBOL(qdisc_put);
993
994/* Version of qdisc_put() that is called with rtnl mutex unlocked.
995 * Intended to be used as optimization, this function only takes rtnl lock if
996 * qdisc reference counter reached zero.
997 */
998
999void qdisc_put_unlocked(struct Qdisc *qdisc)
1000{
1001 if (qdisc->flags & TCQ_F_BUILTIN ||
1002 !refcount_dec_and_rtnl_lock(&qdisc->refcnt))
1003 return;
1004
1005 qdisc_destroy(qdisc);
1006 rtnl_unlock();
1007}
1008EXPORT_SYMBOL(qdisc_put_unlocked);
1009
1010/* Attach toplevel qdisc to device queue. */
1011struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
1012 struct Qdisc *qdisc)
1013{
1014 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
1015 spinlock_t *root_lock;
1016
1017 root_lock = qdisc_lock(oqdisc);
1018 spin_lock_bh(root_lock);
1019
1020 /* ... and graft new one */
1021 if (qdisc == NULL)
1022 qdisc = &noop_qdisc;
1023 dev_queue->qdisc_sleeping = qdisc;
1024 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1025
1026 spin_unlock_bh(root_lock);
1027
1028 return oqdisc;
1029}
1030EXPORT_SYMBOL(dev_graft_qdisc);
1031
1032static void attach_one_default_qdisc(struct net_device *dev,
1033 struct netdev_queue *dev_queue,
1034 void *_unused)
1035{
1036 struct Qdisc *qdisc;
1037 const struct Qdisc_ops *ops = default_qdisc_ops;
1038
1039 if (dev->priv_flags & IFF_NO_QUEUE)
1040 ops = &noqueue_qdisc_ops;
1041 else if(dev->type == ARPHRD_CAN)
1042 ops = &pfifo_fast_ops;
1043
1044 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1045 if (!qdisc) {
1046 netdev_info(dev, "activation failed\n");
1047 return;
1048 }
1049 if (!netif_is_multiqueue(dev))
1050 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1051 dev_queue->qdisc_sleeping = qdisc;
1052}
1053
1054static void attach_default_qdiscs(struct net_device *dev)
1055{
1056 struct netdev_queue *txq;
1057 struct Qdisc *qdisc;
1058
1059 txq = netdev_get_tx_queue(dev, 0);
1060
1061 if (!netif_is_multiqueue(dev) ||
1062 dev->priv_flags & IFF_NO_QUEUE) {
1063 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1064 dev->qdisc = txq->qdisc_sleeping;
1065 qdisc_refcount_inc(dev->qdisc);
1066 } else {
1067 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1068 if (qdisc) {
1069 dev->qdisc = qdisc;
1070 qdisc->ops->attach(qdisc);
1071 }
1072 }
1073#ifdef CONFIG_NET_SCHED
1074 if (dev->qdisc != &noop_qdisc)
1075 qdisc_hash_add(dev->qdisc, false);
1076#endif
1077}
1078
1079static void transition_one_qdisc(struct net_device *dev,
1080 struct netdev_queue *dev_queue,
1081 void *_need_watchdog)
1082{
1083 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1084 int *need_watchdog_p = _need_watchdog;
1085
1086 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1087 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1088
1089 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1090 if (need_watchdog_p) {
1091 dev_queue->trans_start = 0;
1092 *need_watchdog_p = 1;
1093 }
1094}
1095
1096void dev_activate(struct net_device *dev)
1097{
1098 int need_watchdog;
1099
1100 /* No queueing discipline is attached to device;
1101 * create default one for devices, which need queueing
1102 * and noqueue_qdisc for virtual interfaces
1103 */
1104
1105 if (dev->qdisc == &noop_qdisc)
1106 attach_default_qdiscs(dev);
1107
1108 if (!netif_carrier_ok(dev))
1109 /* Delay activation until next carrier-on event */
1110 return;
1111
1112 need_watchdog = 0;
1113 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1114 if (dev_ingress_queue(dev))
1115 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1116
1117 if (need_watchdog) {
1118 netif_trans_update(dev);
1119 dev_watchdog_up(dev);
1120 }
1121}
1122EXPORT_SYMBOL(dev_activate);
1123
1124static void dev_deactivate_queue(struct net_device *dev,
1125 struct netdev_queue *dev_queue,
1126 void *_qdisc_default)
1127{
1128 struct Qdisc *qdisc_default = _qdisc_default;
1129 struct Qdisc *qdisc;
1130
1131 qdisc = rtnl_dereference(dev_queue->qdisc);
1132 if (qdisc) {
1133 bool nolock = qdisc->flags & TCQ_F_NOLOCK;
1134
1135 if (nolock)
1136 spin_lock_bh(&qdisc->seqlock);
1137 spin_lock_bh(qdisc_lock(qdisc));
1138
1139 if (!(qdisc->flags & TCQ_F_BUILTIN))
1140 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1141
1142 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1143 qdisc_reset(qdisc);
1144
1145 spin_unlock_bh(qdisc_lock(qdisc));
1146 if (nolock)
1147 spin_unlock_bh(&qdisc->seqlock);
1148 }
1149}
1150
1151static bool some_qdisc_is_busy(struct net_device *dev)
1152{
1153 unsigned int i;
1154
1155 for (i = 0; i < dev->num_tx_queues; i++) {
1156 struct netdev_queue *dev_queue;
1157 spinlock_t *root_lock;
1158 struct Qdisc *q;
1159 int val;
1160
1161 dev_queue = netdev_get_tx_queue(dev, i);
1162 q = dev_queue->qdisc_sleeping;
1163
1164 root_lock = qdisc_lock(q);
1165 spin_lock_bh(root_lock);
1166
1167 val = (qdisc_is_running(q) ||
1168 test_bit(__QDISC_STATE_SCHED, &q->state));
1169
1170 spin_unlock_bh(root_lock);
1171
1172 if (val)
1173 return true;
1174 }
1175 return false;
1176}
1177
1178static void dev_qdisc_reset(struct net_device *dev,
1179 struct netdev_queue *dev_queue,
1180 void *none)
1181{
1182 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1183
1184 if (qdisc)
1185 qdisc_reset(qdisc);
1186}
1187
1188/**
1189 * dev_deactivate_many - deactivate transmissions on several devices
1190 * @head: list of devices to deactivate
1191 *
1192 * This function returns only when all outstanding transmissions
1193 * have completed, unless all devices are in dismantle phase.
1194 */
1195void dev_deactivate_many(struct list_head *head)
1196{
1197 struct net_device *dev;
1198
1199 list_for_each_entry(dev, head, close_list) {
1200 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1201 &noop_qdisc);
1202 if (dev_ingress_queue(dev))
1203 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1204 &noop_qdisc);
1205
1206 dev_watchdog_down(dev);
1207 }
1208
1209 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1210 * This is avoided if all devices are in dismantle phase :
1211 * Caller will call synchronize_net() for us
1212 */
1213 synchronize_net();
1214
1215 /* Wait for outstanding qdisc_run calls. */
1216 list_for_each_entry(dev, head, close_list) {
1217 while (some_qdisc_is_busy(dev))
1218 yield();
1219 /* The new qdisc is assigned at this point so we can safely
1220 * unwind stale skb lists and qdisc statistics
1221 */
1222 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1223 if (dev_ingress_queue(dev))
1224 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1225 }
1226}
1227
1228void dev_deactivate(struct net_device *dev)
1229{
1230 LIST_HEAD(single);
1231
1232 list_add(&dev->close_list, &single);
1233 dev_deactivate_many(&single);
1234 list_del(&single);
1235}
1236EXPORT_SYMBOL(dev_deactivate);
1237
1238static int qdisc_change_tx_queue_len(struct net_device *dev,
1239 struct netdev_queue *dev_queue)
1240{
1241 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1242 const struct Qdisc_ops *ops = qdisc->ops;
1243
1244 if (ops->change_tx_queue_len)
1245 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1246 return 0;
1247}
1248
1249int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1250{
1251 bool up = dev->flags & IFF_UP;
1252 unsigned int i;
1253 int ret = 0;
1254
1255 if (up)
1256 dev_deactivate(dev);
1257
1258 for (i = 0; i < dev->num_tx_queues; i++) {
1259 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1260
1261 /* TODO: revert changes on a partial failure */
1262 if (ret)
1263 break;
1264 }
1265
1266 if (up)
1267 dev_activate(dev);
1268 return ret;
1269}
1270
1271static void dev_init_scheduler_queue(struct net_device *dev,
1272 struct netdev_queue *dev_queue,
1273 void *_qdisc)
1274{
1275 struct Qdisc *qdisc = _qdisc;
1276
1277 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1278 dev_queue->qdisc_sleeping = qdisc;
1279}
1280
1281void dev_init_scheduler(struct net_device *dev)
1282{
1283 dev->qdisc = &noop_qdisc;
1284 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1287
1288 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1289}
1290
1291static void shutdown_scheduler_queue(struct net_device *dev,
1292 struct netdev_queue *dev_queue,
1293 void *_qdisc_default)
1294{
1295 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1296 struct Qdisc *qdisc_default = _qdisc_default;
1297
1298 if (qdisc) {
1299 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1300 dev_queue->qdisc_sleeping = qdisc_default;
1301
1302 qdisc_put(qdisc);
1303 }
1304}
1305
1306void dev_shutdown(struct net_device *dev)
1307{
1308 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1309 if (dev_ingress_queue(dev))
1310 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1311 qdisc_put(dev->qdisc);
1312 dev->qdisc = &noop_qdisc;
1313
1314 WARN_ON(timer_pending(&dev->watchdog_timer));
1315}
1316
1317void psched_ratecfg_precompute(struct psched_ratecfg *r,
1318 const struct tc_ratespec *conf,
1319 u64 rate64)
1320{
1321 memset(r, 0, sizeof(*r));
1322 r->overhead = conf->overhead;
1323 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1324 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1325 r->mult = 1;
1326 /*
1327 * The deal here is to replace a divide by a reciprocal one
1328 * in fast path (a reciprocal divide is a multiply and a shift)
1329 *
1330 * Normal formula would be :
1331 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1332 *
1333 * We compute mult/shift to use instead :
1334 * time_in_ns = (len * mult) >> shift;
1335 *
1336 * We try to get the highest possible mult value for accuracy,
1337 * but have to make sure no overflows will ever happen.
1338 */
1339 if (r->rate_bytes_ps > 0) {
1340 u64 factor = NSEC_PER_SEC;
1341
1342 for (;;) {
1343 r->mult = div64_u64(factor, r->rate_bytes_ps);
1344 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1345 break;
1346 factor <<= 1;
1347 r->shift++;
1348 }
1349 }
1350}
1351EXPORT_SYMBOL(psched_ratecfg_precompute);
1352
1353static void mini_qdisc_rcu_func(struct rcu_head *head)
1354{
1355}
1356
1357void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1358 struct tcf_proto *tp_head)
1359{
1360 /* Protected with chain0->filter_chain_lock.
1361 * Can't access chain directly because tp_head can be NULL.
1362 */
1363 struct mini_Qdisc *miniq_old =
1364 rcu_dereference_protected(*miniqp->p_miniq, 1);
1365 struct mini_Qdisc *miniq;
1366
1367 if (!tp_head) {
1368 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1369 /* Wait for flying RCU callback before it is freed. */
1370 rcu_barrier();
1371 return;
1372 }
1373
1374 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1375 &miniqp->miniq1 : &miniqp->miniq2;
1376
1377 /* We need to make sure that readers won't see the miniq
1378 * we are about to modify. So wait until previous call_rcu callback
1379 * is done.
1380 */
1381 rcu_barrier();
1382 miniq->filter_list = tp_head;
1383 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1384
1385 if (miniq_old)
1386 /* This is counterpart of the rcu barriers above. We need to
1387 * block potential new user of miniq_old until all readers
1388 * are not seeing it.
1389 */
1390 call_rcu(&miniq_old->rcu, mini_qdisc_rcu_func);
1391}
1392EXPORT_SYMBOL(mini_qdisc_pair_swap);
1393
1394void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1395 struct mini_Qdisc __rcu **p_miniq)
1396{
1397 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1398 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1399 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1400 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1401 miniqp->p_miniq = p_miniq;
1402}
1403EXPORT_SYMBOL(mini_qdisc_pair_init);
1/*
2 * net/sched/sch_generic.c Generic packet scheduler routines.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
8 *
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
10 * Jamal Hadi Salim, <hadi@cyberus.ca> 990601
11 * - Ingress support
12 */
13
14#include <linux/bitops.h>
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/kernel.h>
18#include <linux/sched.h>
19#include <linux/string.h>
20#include <linux/errno.h>
21#include <linux/netdevice.h>
22#include <linux/skbuff.h>
23#include <linux/rtnetlink.h>
24#include <linux/init.h>
25#include <linux/rcupdate.h>
26#include <linux/list.h>
27#include <linux/slab.h>
28#include <linux/if_vlan.h>
29#include <linux/skb_array.h>
30#include <linux/if_macvlan.h>
31#include <net/sch_generic.h>
32#include <net/pkt_sched.h>
33#include <net/dst.h>
34#include <trace/events/qdisc.h>
35#include <net/xfrm.h>
36
37/* Qdisc to use by default */
38const struct Qdisc_ops *default_qdisc_ops = &pfifo_fast_ops;
39EXPORT_SYMBOL(default_qdisc_ops);
40
41/* Main transmission queue. */
42
43/* Modifications to data participating in scheduling must be protected with
44 * qdisc_lock(qdisc) spinlock.
45 *
46 * The idea is the following:
47 * - enqueue, dequeue are serialized via qdisc root lock
48 * - ingress filtering is also serialized via qdisc root lock
49 * - updates to tree and tree walking are only done under the rtnl mutex.
50 */
51
52static inline struct sk_buff *__skb_dequeue_bad_txq(struct Qdisc *q)
53{
54 const struct netdev_queue *txq = q->dev_queue;
55 spinlock_t *lock = NULL;
56 struct sk_buff *skb;
57
58 if (q->flags & TCQ_F_NOLOCK) {
59 lock = qdisc_lock(q);
60 spin_lock(lock);
61 }
62
63 skb = skb_peek(&q->skb_bad_txq);
64 if (skb) {
65 /* check the reason of requeuing without tx lock first */
66 txq = skb_get_tx_queue(txq->dev, skb);
67 if (!netif_xmit_frozen_or_stopped(txq)) {
68 skb = __skb_dequeue(&q->skb_bad_txq);
69 if (qdisc_is_percpu_stats(q)) {
70 qdisc_qstats_cpu_backlog_dec(q, skb);
71 qdisc_qstats_cpu_qlen_dec(q);
72 } else {
73 qdisc_qstats_backlog_dec(q, skb);
74 q->q.qlen--;
75 }
76 } else {
77 skb = NULL;
78 }
79 }
80
81 if (lock)
82 spin_unlock(lock);
83
84 return skb;
85}
86
87static inline struct sk_buff *qdisc_dequeue_skb_bad_txq(struct Qdisc *q)
88{
89 struct sk_buff *skb = skb_peek(&q->skb_bad_txq);
90
91 if (unlikely(skb))
92 skb = __skb_dequeue_bad_txq(q);
93
94 return skb;
95}
96
97static inline void qdisc_enqueue_skb_bad_txq(struct Qdisc *q,
98 struct sk_buff *skb)
99{
100 spinlock_t *lock = NULL;
101
102 if (q->flags & TCQ_F_NOLOCK) {
103 lock = qdisc_lock(q);
104 spin_lock(lock);
105 }
106
107 __skb_queue_tail(&q->skb_bad_txq, skb);
108
109 if (qdisc_is_percpu_stats(q)) {
110 qdisc_qstats_cpu_backlog_inc(q, skb);
111 qdisc_qstats_cpu_qlen_inc(q);
112 } else {
113 qdisc_qstats_backlog_inc(q, skb);
114 q->q.qlen++;
115 }
116
117 if (lock)
118 spin_unlock(lock);
119}
120
121static inline int __dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
122{
123 while (skb) {
124 struct sk_buff *next = skb->next;
125
126 __skb_queue_tail(&q->gso_skb, skb);
127 q->qstats.requeues++;
128 qdisc_qstats_backlog_inc(q, skb);
129 q->q.qlen++; /* it's still part of the queue */
130
131 skb = next;
132 }
133 __netif_schedule(q);
134
135 return 0;
136}
137
138static inline int dev_requeue_skb_locked(struct sk_buff *skb, struct Qdisc *q)
139{
140 spinlock_t *lock = qdisc_lock(q);
141
142 spin_lock(lock);
143 while (skb) {
144 struct sk_buff *next = skb->next;
145
146 __skb_queue_tail(&q->gso_skb, skb);
147
148 qdisc_qstats_cpu_requeues_inc(q);
149 qdisc_qstats_cpu_backlog_inc(q, skb);
150 qdisc_qstats_cpu_qlen_inc(q);
151
152 skb = next;
153 }
154 spin_unlock(lock);
155
156 __netif_schedule(q);
157
158 return 0;
159}
160
161static inline int dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q)
162{
163 if (q->flags & TCQ_F_NOLOCK)
164 return dev_requeue_skb_locked(skb, q);
165 else
166 return __dev_requeue_skb(skb, q);
167}
168
169static void try_bulk_dequeue_skb(struct Qdisc *q,
170 struct sk_buff *skb,
171 const struct netdev_queue *txq,
172 int *packets)
173{
174 int bytelimit = qdisc_avail_bulklimit(txq) - skb->len;
175
176 while (bytelimit > 0) {
177 struct sk_buff *nskb = q->dequeue(q);
178
179 if (!nskb)
180 break;
181
182 bytelimit -= nskb->len; /* covers GSO len */
183 skb->next = nskb;
184 skb = nskb;
185 (*packets)++; /* GSO counts as one pkt */
186 }
187 skb->next = NULL;
188}
189
190/* This variant of try_bulk_dequeue_skb() makes sure
191 * all skbs in the chain are for the same txq
192 */
193static void try_bulk_dequeue_skb_slow(struct Qdisc *q,
194 struct sk_buff *skb,
195 int *packets)
196{
197 int mapping = skb_get_queue_mapping(skb);
198 struct sk_buff *nskb;
199 int cnt = 0;
200
201 do {
202 nskb = q->dequeue(q);
203 if (!nskb)
204 break;
205 if (unlikely(skb_get_queue_mapping(nskb) != mapping)) {
206 qdisc_enqueue_skb_bad_txq(q, nskb);
207 break;
208 }
209 skb->next = nskb;
210 skb = nskb;
211 } while (++cnt < 8);
212 (*packets) += cnt;
213 skb->next = NULL;
214}
215
216/* Note that dequeue_skb can possibly return a SKB list (via skb->next).
217 * A requeued skb (via q->gso_skb) can also be a SKB list.
218 */
219static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate,
220 int *packets)
221{
222 const struct netdev_queue *txq = q->dev_queue;
223 struct sk_buff *skb = NULL;
224
225 *packets = 1;
226 if (unlikely(!skb_queue_empty(&q->gso_skb))) {
227 spinlock_t *lock = NULL;
228
229 if (q->flags & TCQ_F_NOLOCK) {
230 lock = qdisc_lock(q);
231 spin_lock(lock);
232 }
233
234 skb = skb_peek(&q->gso_skb);
235
236 /* skb may be null if another cpu pulls gso_skb off in between
237 * empty check and lock.
238 */
239 if (!skb) {
240 if (lock)
241 spin_unlock(lock);
242 goto validate;
243 }
244
245 /* skb in gso_skb were already validated */
246 *validate = false;
247 if (xfrm_offload(skb))
248 *validate = true;
249 /* check the reason of requeuing without tx lock first */
250 txq = skb_get_tx_queue(txq->dev, skb);
251 if (!netif_xmit_frozen_or_stopped(txq)) {
252 skb = __skb_dequeue(&q->gso_skb);
253 if (qdisc_is_percpu_stats(q)) {
254 qdisc_qstats_cpu_backlog_dec(q, skb);
255 qdisc_qstats_cpu_qlen_dec(q);
256 } else {
257 qdisc_qstats_backlog_dec(q, skb);
258 q->q.qlen--;
259 }
260 } else {
261 skb = NULL;
262 }
263 if (lock)
264 spin_unlock(lock);
265 goto trace;
266 }
267validate:
268 *validate = true;
269
270 if ((q->flags & TCQ_F_ONETXQUEUE) &&
271 netif_xmit_frozen_or_stopped(txq))
272 return skb;
273
274 skb = qdisc_dequeue_skb_bad_txq(q);
275 if (unlikely(skb))
276 goto bulk;
277 skb = q->dequeue(q);
278 if (skb) {
279bulk:
280 if (qdisc_may_bulk(q))
281 try_bulk_dequeue_skb(q, skb, txq, packets);
282 else
283 try_bulk_dequeue_skb_slow(q, skb, packets);
284 }
285trace:
286 trace_qdisc_dequeue(q, txq, *packets, skb);
287 return skb;
288}
289
290/*
291 * Transmit possibly several skbs, and handle the return status as
292 * required. Owning running seqcount bit guarantees that
293 * only one CPU can execute this function.
294 *
295 * Returns to the caller:
296 * false - hardware queue frozen backoff
297 * true - feel free to send more pkts
298 */
299bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q,
300 struct net_device *dev, struct netdev_queue *txq,
301 spinlock_t *root_lock, bool validate)
302{
303 int ret = NETDEV_TX_BUSY;
304 bool again = false;
305
306 /* And release qdisc */
307 if (root_lock)
308 spin_unlock(root_lock);
309
310 /* Note that we validate skb (GSO, checksum, ...) outside of locks */
311 if (validate)
312 skb = validate_xmit_skb_list(skb, dev, &again);
313
314#ifdef CONFIG_XFRM_OFFLOAD
315 if (unlikely(again)) {
316 if (root_lock)
317 spin_lock(root_lock);
318
319 dev_requeue_skb(skb, q);
320 return false;
321 }
322#endif
323
324 if (likely(skb)) {
325 HARD_TX_LOCK(dev, txq, smp_processor_id());
326 if (!netif_xmit_frozen_or_stopped(txq))
327 skb = dev_hard_start_xmit(skb, dev, txq, &ret);
328
329 HARD_TX_UNLOCK(dev, txq);
330 } else {
331 if (root_lock)
332 spin_lock(root_lock);
333 return true;
334 }
335
336 if (root_lock)
337 spin_lock(root_lock);
338
339 if (!dev_xmit_complete(ret)) {
340 /* Driver returned NETDEV_TX_BUSY - requeue skb */
341 if (unlikely(ret != NETDEV_TX_BUSY))
342 net_warn_ratelimited("BUG %s code %d qlen %d\n",
343 dev->name, ret, q->q.qlen);
344
345 dev_requeue_skb(skb, q);
346 return false;
347 }
348
349 if (ret && netif_xmit_frozen_or_stopped(txq))
350 return false;
351
352 return true;
353}
354
355/*
356 * NOTE: Called under qdisc_lock(q) with locally disabled BH.
357 *
358 * running seqcount guarantees only one CPU can process
359 * this qdisc at a time. qdisc_lock(q) serializes queue accesses for
360 * this queue.
361 *
362 * netif_tx_lock serializes accesses to device driver.
363 *
364 * qdisc_lock(q) and netif_tx_lock are mutually exclusive,
365 * if one is grabbed, another must be free.
366 *
367 * Note, that this procedure can be called by a watchdog timer
368 *
369 * Returns to the caller:
370 * 0 - queue is empty or throttled.
371 * >0 - queue is not empty.
372 *
373 */
374static inline bool qdisc_restart(struct Qdisc *q, int *packets)
375{
376 bool more, validate, nolock = q->flags & TCQ_F_NOLOCK;
377 spinlock_t *root_lock = NULL;
378 struct netdev_queue *txq;
379 struct net_device *dev;
380 struct sk_buff *skb;
381
382 /* Dequeue packet */
383 if (nolock && test_and_set_bit(__QDISC_STATE_RUNNING, &q->state))
384 return false;
385
386 skb = dequeue_skb(q, &validate, packets);
387 if (unlikely(!skb)) {
388 if (nolock)
389 clear_bit(__QDISC_STATE_RUNNING, &q->state);
390 return false;
391 }
392
393 if (!nolock)
394 root_lock = qdisc_lock(q);
395
396 dev = qdisc_dev(q);
397 txq = skb_get_tx_queue(dev, skb);
398
399 more = sch_direct_xmit(skb, q, dev, txq, root_lock, validate);
400 if (nolock)
401 clear_bit(__QDISC_STATE_RUNNING, &q->state);
402 return more;
403}
404
405void __qdisc_run(struct Qdisc *q)
406{
407 int quota = dev_tx_weight;
408 int packets;
409
410 while (qdisc_restart(q, &packets)) {
411 /*
412 * Ordered by possible occurrence: Postpone processing if
413 * 1. we've exceeded packet quota
414 * 2. another process needs the CPU;
415 */
416 quota -= packets;
417 if (quota <= 0 || need_resched()) {
418 __netif_schedule(q);
419 break;
420 }
421 }
422}
423
424unsigned long dev_trans_start(struct net_device *dev)
425{
426 unsigned long val, res;
427 unsigned int i;
428
429 if (is_vlan_dev(dev))
430 dev = vlan_dev_real_dev(dev);
431 else if (netif_is_macvlan(dev))
432 dev = macvlan_dev_real_dev(dev);
433 res = netdev_get_tx_queue(dev, 0)->trans_start;
434 for (i = 1; i < dev->num_tx_queues; i++) {
435 val = netdev_get_tx_queue(dev, i)->trans_start;
436 if (val && time_after(val, res))
437 res = val;
438 }
439
440 return res;
441}
442EXPORT_SYMBOL(dev_trans_start);
443
444static void dev_watchdog(struct timer_list *t)
445{
446 struct net_device *dev = from_timer(dev, t, watchdog_timer);
447
448 netif_tx_lock(dev);
449 if (!qdisc_tx_is_noop(dev)) {
450 if (netif_device_present(dev) &&
451 netif_running(dev) &&
452 netif_carrier_ok(dev)) {
453 int some_queue_timedout = 0;
454 unsigned int i;
455 unsigned long trans_start;
456
457 for (i = 0; i < dev->num_tx_queues; i++) {
458 struct netdev_queue *txq;
459
460 txq = netdev_get_tx_queue(dev, i);
461 trans_start = txq->trans_start;
462 if (netif_xmit_stopped(txq) &&
463 time_after(jiffies, (trans_start +
464 dev->watchdog_timeo))) {
465 some_queue_timedout = 1;
466 txq->trans_timeout++;
467 break;
468 }
469 }
470
471 if (some_queue_timedout) {
472 WARN_ONCE(1, KERN_INFO "NETDEV WATCHDOG: %s (%s): transmit queue %u timed out\n",
473 dev->name, netdev_drivername(dev), i);
474 dev->netdev_ops->ndo_tx_timeout(dev);
475 }
476 if (!mod_timer(&dev->watchdog_timer,
477 round_jiffies(jiffies +
478 dev->watchdog_timeo)))
479 dev_hold(dev);
480 }
481 }
482 netif_tx_unlock(dev);
483
484 dev_put(dev);
485}
486
487void __netdev_watchdog_up(struct net_device *dev)
488{
489 if (dev->netdev_ops->ndo_tx_timeout) {
490 if (dev->watchdog_timeo <= 0)
491 dev->watchdog_timeo = 5*HZ;
492 if (!mod_timer(&dev->watchdog_timer,
493 round_jiffies(jiffies + dev->watchdog_timeo)))
494 dev_hold(dev);
495 }
496}
497
498static void dev_watchdog_up(struct net_device *dev)
499{
500 __netdev_watchdog_up(dev);
501}
502
503static void dev_watchdog_down(struct net_device *dev)
504{
505 netif_tx_lock_bh(dev);
506 if (del_timer(&dev->watchdog_timer))
507 dev_put(dev);
508 netif_tx_unlock_bh(dev);
509}
510
511/**
512 * netif_carrier_on - set carrier
513 * @dev: network device
514 *
515 * Device has detected that carrier.
516 */
517void netif_carrier_on(struct net_device *dev)
518{
519 if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
520 if (dev->reg_state == NETREG_UNINITIALIZED)
521 return;
522 atomic_inc(&dev->carrier_up_count);
523 linkwatch_fire_event(dev);
524 if (netif_running(dev))
525 __netdev_watchdog_up(dev);
526 }
527}
528EXPORT_SYMBOL(netif_carrier_on);
529
530/**
531 * netif_carrier_off - clear carrier
532 * @dev: network device
533 *
534 * Device has detected loss of carrier.
535 */
536void netif_carrier_off(struct net_device *dev)
537{
538 if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
539 if (dev->reg_state == NETREG_UNINITIALIZED)
540 return;
541 atomic_inc(&dev->carrier_down_count);
542 linkwatch_fire_event(dev);
543 }
544}
545EXPORT_SYMBOL(netif_carrier_off);
546
547/* "NOOP" scheduler: the best scheduler, recommended for all interfaces
548 under all circumstances. It is difficult to invent anything faster or
549 cheaper.
550 */
551
552static int noop_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
553 struct sk_buff **to_free)
554{
555 __qdisc_drop(skb, to_free);
556 return NET_XMIT_CN;
557}
558
559static struct sk_buff *noop_dequeue(struct Qdisc *qdisc)
560{
561 return NULL;
562}
563
564struct Qdisc_ops noop_qdisc_ops __read_mostly = {
565 .id = "noop",
566 .priv_size = 0,
567 .enqueue = noop_enqueue,
568 .dequeue = noop_dequeue,
569 .peek = noop_dequeue,
570 .owner = THIS_MODULE,
571};
572
573static struct netdev_queue noop_netdev_queue = {
574 .qdisc = &noop_qdisc,
575 .qdisc_sleeping = &noop_qdisc,
576};
577
578struct Qdisc noop_qdisc = {
579 .enqueue = noop_enqueue,
580 .dequeue = noop_dequeue,
581 .flags = TCQ_F_BUILTIN,
582 .ops = &noop_qdisc_ops,
583 .q.lock = __SPIN_LOCK_UNLOCKED(noop_qdisc.q.lock),
584 .dev_queue = &noop_netdev_queue,
585 .running = SEQCNT_ZERO(noop_qdisc.running),
586 .busylock = __SPIN_LOCK_UNLOCKED(noop_qdisc.busylock),
587};
588EXPORT_SYMBOL(noop_qdisc);
589
590static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt,
591 struct netlink_ext_ack *extack)
592{
593 /* register_qdisc() assigns a default of noop_enqueue if unset,
594 * but __dev_queue_xmit() treats noqueue only as such
595 * if this is NULL - so clear it here. */
596 qdisc->enqueue = NULL;
597 return 0;
598}
599
600struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {
601 .id = "noqueue",
602 .priv_size = 0,
603 .init = noqueue_init,
604 .enqueue = noop_enqueue,
605 .dequeue = noop_dequeue,
606 .peek = noop_dequeue,
607 .owner = THIS_MODULE,
608};
609
610static const u8 prio2band[TC_PRIO_MAX + 1] = {
611 1, 2, 2, 2, 1, 2, 0, 0 , 1, 1, 1, 1, 1, 1, 1, 1
612};
613
614/* 3-band FIFO queue: old style, but should be a bit faster than
615 generic prio+fifo combination.
616 */
617
618#define PFIFO_FAST_BANDS 3
619
620/*
621 * Private data for a pfifo_fast scheduler containing:
622 * - rings for priority bands
623 */
624struct pfifo_fast_priv {
625 struct skb_array q[PFIFO_FAST_BANDS];
626};
627
628static inline struct skb_array *band2list(struct pfifo_fast_priv *priv,
629 int band)
630{
631 return &priv->q[band];
632}
633
634static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc *qdisc,
635 struct sk_buff **to_free)
636{
637 int band = prio2band[skb->priority & TC_PRIO_MAX];
638 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
639 struct skb_array *q = band2list(priv, band);
640 unsigned int pkt_len = qdisc_pkt_len(skb);
641 int err;
642
643 err = skb_array_produce(q, skb);
644
645 if (unlikely(err))
646 return qdisc_drop_cpu(skb, qdisc, to_free);
647
648 qdisc_qstats_cpu_qlen_inc(qdisc);
649 /* Note: skb can not be used after skb_array_produce(),
650 * so we better not use qdisc_qstats_cpu_backlog_inc()
651 */
652 this_cpu_add(qdisc->cpu_qstats->backlog, pkt_len);
653 return NET_XMIT_SUCCESS;
654}
655
656static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc)
657{
658 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
659 struct sk_buff *skb = NULL;
660 int band;
661
662 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
663 struct skb_array *q = band2list(priv, band);
664
665 if (__skb_array_empty(q))
666 continue;
667
668 skb = skb_array_consume_bh(q);
669 }
670 if (likely(skb)) {
671 qdisc_qstats_cpu_backlog_dec(qdisc, skb);
672 qdisc_bstats_cpu_update(qdisc, skb);
673 qdisc_qstats_cpu_qlen_dec(qdisc);
674 }
675
676 return skb;
677}
678
679static struct sk_buff *pfifo_fast_peek(struct Qdisc *qdisc)
680{
681 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
682 struct sk_buff *skb = NULL;
683 int band;
684
685 for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) {
686 struct skb_array *q = band2list(priv, band);
687
688 skb = __skb_array_peek(q);
689 }
690
691 return skb;
692}
693
694static void pfifo_fast_reset(struct Qdisc *qdisc)
695{
696 int i, band;
697 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
698
699 for (band = 0; band < PFIFO_FAST_BANDS; band++) {
700 struct skb_array *q = band2list(priv, band);
701 struct sk_buff *skb;
702
703 /* NULL ring is possible if destroy path is due to a failed
704 * skb_array_init() in pfifo_fast_init() case.
705 */
706 if (!q->ring.queue)
707 continue;
708
709 while ((skb = skb_array_consume_bh(q)) != NULL)
710 kfree_skb(skb);
711 }
712
713 for_each_possible_cpu(i) {
714 struct gnet_stats_queue *q = per_cpu_ptr(qdisc->cpu_qstats, i);
715
716 q->backlog = 0;
717 q->qlen = 0;
718 }
719}
720
721static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
722{
723 struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
724
725 memcpy(&opt.priomap, prio2band, TC_PRIO_MAX + 1);
726 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
727 goto nla_put_failure;
728 return skb->len;
729
730nla_put_failure:
731 return -1;
732}
733
734static int pfifo_fast_init(struct Qdisc *qdisc, struct nlattr *opt,
735 struct netlink_ext_ack *extack)
736{
737 unsigned int qlen = qdisc_dev(qdisc)->tx_queue_len;
738 struct pfifo_fast_priv *priv = qdisc_priv(qdisc);
739 int prio;
740
741 /* guard against zero length rings */
742 if (!qlen)
743 return -EINVAL;
744
745 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
746 struct skb_array *q = band2list(priv, prio);
747 int err;
748
749 err = skb_array_init(q, qlen, GFP_KERNEL);
750 if (err)
751 return -ENOMEM;
752 }
753
754 /* Can by-pass the queue discipline */
755 qdisc->flags |= TCQ_F_CAN_BYPASS;
756 return 0;
757}
758
759static void pfifo_fast_destroy(struct Qdisc *sch)
760{
761 struct pfifo_fast_priv *priv = qdisc_priv(sch);
762 int prio;
763
764 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
765 struct skb_array *q = band2list(priv, prio);
766
767 /* NULL ring is possible if destroy path is due to a failed
768 * skb_array_init() in pfifo_fast_init() case.
769 */
770 if (!q->ring.queue)
771 continue;
772 /* Destroy ring but no need to kfree_skb because a call to
773 * pfifo_fast_reset() has already done that work.
774 */
775 ptr_ring_cleanup(&q->ring, NULL);
776 }
777}
778
779static int pfifo_fast_change_tx_queue_len(struct Qdisc *sch,
780 unsigned int new_len)
781{
782 struct pfifo_fast_priv *priv = qdisc_priv(sch);
783 struct skb_array *bands[PFIFO_FAST_BANDS];
784 int prio;
785
786 for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
787 struct skb_array *q = band2list(priv, prio);
788
789 bands[prio] = q;
790 }
791
792 return skb_array_resize_multiple(bands, PFIFO_FAST_BANDS, new_len,
793 GFP_KERNEL);
794}
795
796struct Qdisc_ops pfifo_fast_ops __read_mostly = {
797 .id = "pfifo_fast",
798 .priv_size = sizeof(struct pfifo_fast_priv),
799 .enqueue = pfifo_fast_enqueue,
800 .dequeue = pfifo_fast_dequeue,
801 .peek = pfifo_fast_peek,
802 .init = pfifo_fast_init,
803 .destroy = pfifo_fast_destroy,
804 .reset = pfifo_fast_reset,
805 .dump = pfifo_fast_dump,
806 .change_tx_queue_len = pfifo_fast_change_tx_queue_len,
807 .owner = THIS_MODULE,
808 .static_flags = TCQ_F_NOLOCK | TCQ_F_CPUSTATS,
809};
810EXPORT_SYMBOL(pfifo_fast_ops);
811
812static struct lock_class_key qdisc_tx_busylock;
813static struct lock_class_key qdisc_running_key;
814
815struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
816 const struct Qdisc_ops *ops,
817 struct netlink_ext_ack *extack)
818{
819 void *p;
820 struct Qdisc *sch;
821 unsigned int size = QDISC_ALIGN(sizeof(*sch)) + ops->priv_size;
822 int err = -ENOBUFS;
823 struct net_device *dev;
824
825 if (!dev_queue) {
826 NL_SET_ERR_MSG(extack, "No device queue given");
827 err = -EINVAL;
828 goto errout;
829 }
830
831 dev = dev_queue->dev;
832 p = kzalloc_node(size, GFP_KERNEL,
833 netdev_queue_numa_node_read(dev_queue));
834
835 if (!p)
836 goto errout;
837 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
838 /* if we got non aligned memory, ask more and do alignment ourself */
839 if (sch != p) {
840 kfree(p);
841 p = kzalloc_node(size + QDISC_ALIGNTO - 1, GFP_KERNEL,
842 netdev_queue_numa_node_read(dev_queue));
843 if (!p)
844 goto errout;
845 sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
846 sch->padded = (char *) sch - (char *) p;
847 }
848 __skb_queue_head_init(&sch->gso_skb);
849 __skb_queue_head_init(&sch->skb_bad_txq);
850 qdisc_skb_head_init(&sch->q);
851 spin_lock_init(&sch->q.lock);
852
853 if (ops->static_flags & TCQ_F_CPUSTATS) {
854 sch->cpu_bstats =
855 netdev_alloc_pcpu_stats(struct gnet_stats_basic_cpu);
856 if (!sch->cpu_bstats)
857 goto errout1;
858
859 sch->cpu_qstats = alloc_percpu(struct gnet_stats_queue);
860 if (!sch->cpu_qstats) {
861 free_percpu(sch->cpu_bstats);
862 goto errout1;
863 }
864 }
865
866 spin_lock_init(&sch->busylock);
867 lockdep_set_class(&sch->busylock,
868 dev->qdisc_tx_busylock ?: &qdisc_tx_busylock);
869
870 seqcount_init(&sch->running);
871 lockdep_set_class(&sch->running,
872 dev->qdisc_running_key ?: &qdisc_running_key);
873
874 sch->ops = ops;
875 sch->flags = ops->static_flags;
876 sch->enqueue = ops->enqueue;
877 sch->dequeue = ops->dequeue;
878 sch->dev_queue = dev_queue;
879 dev_hold(dev);
880 refcount_set(&sch->refcnt, 1);
881
882 return sch;
883errout1:
884 kfree(p);
885errout:
886 return ERR_PTR(err);
887}
888
889struct Qdisc *qdisc_create_dflt(struct netdev_queue *dev_queue,
890 const struct Qdisc_ops *ops,
891 unsigned int parentid,
892 struct netlink_ext_ack *extack)
893{
894 struct Qdisc *sch;
895
896 if (!try_module_get(ops->owner)) {
897 NL_SET_ERR_MSG(extack, "Failed to increase module reference counter");
898 return NULL;
899 }
900
901 sch = qdisc_alloc(dev_queue, ops, extack);
902 if (IS_ERR(sch)) {
903 module_put(ops->owner);
904 return NULL;
905 }
906 sch->parent = parentid;
907
908 if (!ops->init || ops->init(sch, NULL, extack) == 0)
909 return sch;
910
911 qdisc_destroy(sch);
912 return NULL;
913}
914EXPORT_SYMBOL(qdisc_create_dflt);
915
916/* Under qdisc_lock(qdisc) and BH! */
917
918void qdisc_reset(struct Qdisc *qdisc)
919{
920 const struct Qdisc_ops *ops = qdisc->ops;
921 struct sk_buff *skb, *tmp;
922
923 if (ops->reset)
924 ops->reset(qdisc);
925
926 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
927 __skb_unlink(skb, &qdisc->gso_skb);
928 kfree_skb_list(skb);
929 }
930
931 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
932 __skb_unlink(skb, &qdisc->skb_bad_txq);
933 kfree_skb_list(skb);
934 }
935
936 qdisc->q.qlen = 0;
937 qdisc->qstats.backlog = 0;
938}
939EXPORT_SYMBOL(qdisc_reset);
940
941void qdisc_free(struct Qdisc *qdisc)
942{
943 if (qdisc_is_percpu_stats(qdisc)) {
944 free_percpu(qdisc->cpu_bstats);
945 free_percpu(qdisc->cpu_qstats);
946 }
947
948 kfree((char *) qdisc - qdisc->padded);
949}
950
951void qdisc_destroy(struct Qdisc *qdisc)
952{
953 const struct Qdisc_ops *ops = qdisc->ops;
954 struct sk_buff *skb, *tmp;
955
956 if (qdisc->flags & TCQ_F_BUILTIN ||
957 !refcount_dec_and_test(&qdisc->refcnt))
958 return;
959
960#ifdef CONFIG_NET_SCHED
961 qdisc_hash_del(qdisc);
962
963 qdisc_put_stab(rtnl_dereference(qdisc->stab));
964#endif
965 gen_kill_estimator(&qdisc->rate_est);
966 if (ops->reset)
967 ops->reset(qdisc);
968 if (ops->destroy)
969 ops->destroy(qdisc);
970
971 module_put(ops->owner);
972 dev_put(qdisc_dev(qdisc));
973
974 skb_queue_walk_safe(&qdisc->gso_skb, skb, tmp) {
975 __skb_unlink(skb, &qdisc->gso_skb);
976 kfree_skb_list(skb);
977 }
978
979 skb_queue_walk_safe(&qdisc->skb_bad_txq, skb, tmp) {
980 __skb_unlink(skb, &qdisc->skb_bad_txq);
981 kfree_skb_list(skb);
982 }
983
984 qdisc_free(qdisc);
985}
986EXPORT_SYMBOL(qdisc_destroy);
987
988/* Attach toplevel qdisc to device queue. */
989struct Qdisc *dev_graft_qdisc(struct netdev_queue *dev_queue,
990 struct Qdisc *qdisc)
991{
992 struct Qdisc *oqdisc = dev_queue->qdisc_sleeping;
993 spinlock_t *root_lock;
994
995 root_lock = qdisc_lock(oqdisc);
996 spin_lock_bh(root_lock);
997
998 /* ... and graft new one */
999 if (qdisc == NULL)
1000 qdisc = &noop_qdisc;
1001 dev_queue->qdisc_sleeping = qdisc;
1002 rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
1003
1004 spin_unlock_bh(root_lock);
1005
1006 return oqdisc;
1007}
1008EXPORT_SYMBOL(dev_graft_qdisc);
1009
1010static void attach_one_default_qdisc(struct net_device *dev,
1011 struct netdev_queue *dev_queue,
1012 void *_unused)
1013{
1014 struct Qdisc *qdisc;
1015 const struct Qdisc_ops *ops = default_qdisc_ops;
1016
1017 if (dev->priv_flags & IFF_NO_QUEUE)
1018 ops = &noqueue_qdisc_ops;
1019
1020 qdisc = qdisc_create_dflt(dev_queue, ops, TC_H_ROOT, NULL);
1021 if (!qdisc) {
1022 netdev_info(dev, "activation failed\n");
1023 return;
1024 }
1025 if (!netif_is_multiqueue(dev))
1026 qdisc->flags |= TCQ_F_ONETXQUEUE | TCQ_F_NOPARENT;
1027 dev_queue->qdisc_sleeping = qdisc;
1028}
1029
1030static void attach_default_qdiscs(struct net_device *dev)
1031{
1032 struct netdev_queue *txq;
1033 struct Qdisc *qdisc;
1034
1035 txq = netdev_get_tx_queue(dev, 0);
1036
1037 if (!netif_is_multiqueue(dev) ||
1038 dev->priv_flags & IFF_NO_QUEUE) {
1039 netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
1040 dev->qdisc = txq->qdisc_sleeping;
1041 qdisc_refcount_inc(dev->qdisc);
1042 } else {
1043 qdisc = qdisc_create_dflt(txq, &mq_qdisc_ops, TC_H_ROOT, NULL);
1044 if (qdisc) {
1045 dev->qdisc = qdisc;
1046 qdisc->ops->attach(qdisc);
1047 }
1048 }
1049#ifdef CONFIG_NET_SCHED
1050 if (dev->qdisc != &noop_qdisc)
1051 qdisc_hash_add(dev->qdisc, false);
1052#endif
1053}
1054
1055static void transition_one_qdisc(struct net_device *dev,
1056 struct netdev_queue *dev_queue,
1057 void *_need_watchdog)
1058{
1059 struct Qdisc *new_qdisc = dev_queue->qdisc_sleeping;
1060 int *need_watchdog_p = _need_watchdog;
1061
1062 if (!(new_qdisc->flags & TCQ_F_BUILTIN))
1063 clear_bit(__QDISC_STATE_DEACTIVATED, &new_qdisc->state);
1064
1065 rcu_assign_pointer(dev_queue->qdisc, new_qdisc);
1066 if (need_watchdog_p) {
1067 dev_queue->trans_start = 0;
1068 *need_watchdog_p = 1;
1069 }
1070}
1071
1072void dev_activate(struct net_device *dev)
1073{
1074 int need_watchdog;
1075
1076 /* No queueing discipline is attached to device;
1077 * create default one for devices, which need queueing
1078 * and noqueue_qdisc for virtual interfaces
1079 */
1080
1081 if (dev->qdisc == &noop_qdisc)
1082 attach_default_qdiscs(dev);
1083
1084 if (!netif_carrier_ok(dev))
1085 /* Delay activation until next carrier-on event */
1086 return;
1087
1088 need_watchdog = 0;
1089 netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
1090 if (dev_ingress_queue(dev))
1091 transition_one_qdisc(dev, dev_ingress_queue(dev), NULL);
1092
1093 if (need_watchdog) {
1094 netif_trans_update(dev);
1095 dev_watchdog_up(dev);
1096 }
1097}
1098EXPORT_SYMBOL(dev_activate);
1099
1100static void dev_deactivate_queue(struct net_device *dev,
1101 struct netdev_queue *dev_queue,
1102 void *_qdisc_default)
1103{
1104 struct Qdisc *qdisc_default = _qdisc_default;
1105 struct Qdisc *qdisc;
1106
1107 qdisc = rtnl_dereference(dev_queue->qdisc);
1108 if (qdisc) {
1109 spin_lock_bh(qdisc_lock(qdisc));
1110
1111 if (!(qdisc->flags & TCQ_F_BUILTIN))
1112 set_bit(__QDISC_STATE_DEACTIVATED, &qdisc->state);
1113
1114 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1115 qdisc_reset(qdisc);
1116
1117 spin_unlock_bh(qdisc_lock(qdisc));
1118 }
1119}
1120
1121static bool some_qdisc_is_busy(struct net_device *dev)
1122{
1123 unsigned int i;
1124
1125 for (i = 0; i < dev->num_tx_queues; i++) {
1126 struct netdev_queue *dev_queue;
1127 spinlock_t *root_lock;
1128 struct Qdisc *q;
1129 int val;
1130
1131 dev_queue = netdev_get_tx_queue(dev, i);
1132 q = dev_queue->qdisc_sleeping;
1133
1134 if (q->flags & TCQ_F_NOLOCK) {
1135 val = test_bit(__QDISC_STATE_SCHED, &q->state);
1136 } else {
1137 root_lock = qdisc_lock(q);
1138 spin_lock_bh(root_lock);
1139
1140 val = (qdisc_is_running(q) ||
1141 test_bit(__QDISC_STATE_SCHED, &q->state));
1142
1143 spin_unlock_bh(root_lock);
1144 }
1145
1146 if (val)
1147 return true;
1148 }
1149 return false;
1150}
1151
1152static void dev_qdisc_reset(struct net_device *dev,
1153 struct netdev_queue *dev_queue,
1154 void *none)
1155{
1156 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1157
1158 if (qdisc)
1159 qdisc_reset(qdisc);
1160}
1161
1162/**
1163 * dev_deactivate_many - deactivate transmissions on several devices
1164 * @head: list of devices to deactivate
1165 *
1166 * This function returns only when all outstanding transmissions
1167 * have completed, unless all devices are in dismantle phase.
1168 */
1169void dev_deactivate_many(struct list_head *head)
1170{
1171 struct net_device *dev;
1172
1173 list_for_each_entry(dev, head, close_list) {
1174 netdev_for_each_tx_queue(dev, dev_deactivate_queue,
1175 &noop_qdisc);
1176 if (dev_ingress_queue(dev))
1177 dev_deactivate_queue(dev, dev_ingress_queue(dev),
1178 &noop_qdisc);
1179
1180 dev_watchdog_down(dev);
1181 }
1182
1183 /* Wait for outstanding qdisc-less dev_queue_xmit calls.
1184 * This is avoided if all devices are in dismantle phase :
1185 * Caller will call synchronize_net() for us
1186 */
1187 synchronize_net();
1188
1189 /* Wait for outstanding qdisc_run calls. */
1190 list_for_each_entry(dev, head, close_list) {
1191 while (some_qdisc_is_busy(dev))
1192 yield();
1193 /* The new qdisc is assigned at this point so we can safely
1194 * unwind stale skb lists and qdisc statistics
1195 */
1196 netdev_for_each_tx_queue(dev, dev_qdisc_reset, NULL);
1197 if (dev_ingress_queue(dev))
1198 dev_qdisc_reset(dev, dev_ingress_queue(dev), NULL);
1199 }
1200}
1201
1202void dev_deactivate(struct net_device *dev)
1203{
1204 LIST_HEAD(single);
1205
1206 list_add(&dev->close_list, &single);
1207 dev_deactivate_many(&single);
1208 list_del(&single);
1209}
1210EXPORT_SYMBOL(dev_deactivate);
1211
1212static int qdisc_change_tx_queue_len(struct net_device *dev,
1213 struct netdev_queue *dev_queue)
1214{
1215 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1216 const struct Qdisc_ops *ops = qdisc->ops;
1217
1218 if (ops->change_tx_queue_len)
1219 return ops->change_tx_queue_len(qdisc, dev->tx_queue_len);
1220 return 0;
1221}
1222
1223int dev_qdisc_change_tx_queue_len(struct net_device *dev)
1224{
1225 bool up = dev->flags & IFF_UP;
1226 unsigned int i;
1227 int ret = 0;
1228
1229 if (up)
1230 dev_deactivate(dev);
1231
1232 for (i = 0; i < dev->num_tx_queues; i++) {
1233 ret = qdisc_change_tx_queue_len(dev, &dev->_tx[i]);
1234
1235 /* TODO: revert changes on a partial failure */
1236 if (ret)
1237 break;
1238 }
1239
1240 if (up)
1241 dev_activate(dev);
1242 return ret;
1243}
1244
1245static void dev_init_scheduler_queue(struct net_device *dev,
1246 struct netdev_queue *dev_queue,
1247 void *_qdisc)
1248{
1249 struct Qdisc *qdisc = _qdisc;
1250
1251 rcu_assign_pointer(dev_queue->qdisc, qdisc);
1252 dev_queue->qdisc_sleeping = qdisc;
1253 __skb_queue_head_init(&qdisc->gso_skb);
1254 __skb_queue_head_init(&qdisc->skb_bad_txq);
1255}
1256
1257void dev_init_scheduler(struct net_device *dev)
1258{
1259 dev->qdisc = &noop_qdisc;
1260 netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
1261 if (dev_ingress_queue(dev))
1262 dev_init_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1263
1264 timer_setup(&dev->watchdog_timer, dev_watchdog, 0);
1265}
1266
1267static void shutdown_scheduler_queue(struct net_device *dev,
1268 struct netdev_queue *dev_queue,
1269 void *_qdisc_default)
1270{
1271 struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
1272 struct Qdisc *qdisc_default = _qdisc_default;
1273
1274 if (qdisc) {
1275 rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
1276 dev_queue->qdisc_sleeping = qdisc_default;
1277
1278 qdisc_destroy(qdisc);
1279 }
1280}
1281
1282void dev_shutdown(struct net_device *dev)
1283{
1284 netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
1285 if (dev_ingress_queue(dev))
1286 shutdown_scheduler_queue(dev, dev_ingress_queue(dev), &noop_qdisc);
1287 qdisc_destroy(dev->qdisc);
1288 dev->qdisc = &noop_qdisc;
1289
1290 WARN_ON(timer_pending(&dev->watchdog_timer));
1291}
1292
1293void psched_ratecfg_precompute(struct psched_ratecfg *r,
1294 const struct tc_ratespec *conf,
1295 u64 rate64)
1296{
1297 memset(r, 0, sizeof(*r));
1298 r->overhead = conf->overhead;
1299 r->rate_bytes_ps = max_t(u64, conf->rate, rate64);
1300 r->linklayer = (conf->linklayer & TC_LINKLAYER_MASK);
1301 r->mult = 1;
1302 /*
1303 * The deal here is to replace a divide by a reciprocal one
1304 * in fast path (a reciprocal divide is a multiply and a shift)
1305 *
1306 * Normal formula would be :
1307 * time_in_ns = (NSEC_PER_SEC * len) / rate_bps
1308 *
1309 * We compute mult/shift to use instead :
1310 * time_in_ns = (len * mult) >> shift;
1311 *
1312 * We try to get the highest possible mult value for accuracy,
1313 * but have to make sure no overflows will ever happen.
1314 */
1315 if (r->rate_bytes_ps > 0) {
1316 u64 factor = NSEC_PER_SEC;
1317
1318 for (;;) {
1319 r->mult = div64_u64(factor, r->rate_bytes_ps);
1320 if (r->mult & (1U << 31) || factor & (1ULL << 63))
1321 break;
1322 factor <<= 1;
1323 r->shift++;
1324 }
1325 }
1326}
1327EXPORT_SYMBOL(psched_ratecfg_precompute);
1328
1329static void mini_qdisc_rcu_func(struct rcu_head *head)
1330{
1331}
1332
1333void mini_qdisc_pair_swap(struct mini_Qdisc_pair *miniqp,
1334 struct tcf_proto *tp_head)
1335{
1336 struct mini_Qdisc *miniq_old = rtnl_dereference(*miniqp->p_miniq);
1337 struct mini_Qdisc *miniq;
1338
1339 if (!tp_head) {
1340 RCU_INIT_POINTER(*miniqp->p_miniq, NULL);
1341 /* Wait for flying RCU callback before it is freed. */
1342 rcu_barrier_bh();
1343 return;
1344 }
1345
1346 miniq = !miniq_old || miniq_old == &miniqp->miniq2 ?
1347 &miniqp->miniq1 : &miniqp->miniq2;
1348
1349 /* We need to make sure that readers won't see the miniq
1350 * we are about to modify. So wait until previous call_rcu_bh callback
1351 * is done.
1352 */
1353 rcu_barrier_bh();
1354 miniq->filter_list = tp_head;
1355 rcu_assign_pointer(*miniqp->p_miniq, miniq);
1356
1357 if (miniq_old)
1358 /* This is counterpart of the rcu barriers above. We need to
1359 * block potential new user of miniq_old until all readers
1360 * are not seeing it.
1361 */
1362 call_rcu_bh(&miniq_old->rcu, mini_qdisc_rcu_func);
1363}
1364EXPORT_SYMBOL(mini_qdisc_pair_swap);
1365
1366void mini_qdisc_pair_init(struct mini_Qdisc_pair *miniqp, struct Qdisc *qdisc,
1367 struct mini_Qdisc __rcu **p_miniq)
1368{
1369 miniqp->miniq1.cpu_bstats = qdisc->cpu_bstats;
1370 miniqp->miniq1.cpu_qstats = qdisc->cpu_qstats;
1371 miniqp->miniq2.cpu_bstats = qdisc->cpu_bstats;
1372 miniqp->miniq2.cpu_qstats = qdisc->cpu_qstats;
1373 miniqp->p_miniq = p_miniq;
1374}
1375EXPORT_SYMBOL(mini_qdisc_pair_init);